[Beowulf] Full-duplex performance problems with Intel Gigabit Ethernet cards?

[Felix Rauch Valenti]

We have a strange performance problem in one of our clusters. It looks
as if some Gigabit Ethernet adapters are running only half-duplex, but
according to the switch and "ethtool" they are running full duplex.

The machines are equipped with Intel Corp. 82541GI/PI (and  82547GI)
Gigabit Ethernet Controllers and 3.2 GHz Pentium4 CPUs. The strange
thing is that when the machines either send OR receive a high-speed
TCP data stream, we see regular performance, but when they have to
send AND receive at the same time, we see only about half the expected
performance. On some older nodes, everything is fine.

More specifically, we have TCP data stream from node A to node B,
which then concurrently forwards the data stream further on to node C:
Node A -> Node B -> Node C.

The performance numbers for the stream are as follows:
Node A -> Node B: 98 MByte/s
Node A -> Node B -> Node C: 50 MByte/s
(In the first case, Node A is sending only and Node B is receiving
only. In the second case, Node B is sending and receiving at the same
time)

This is very strange, especially since the same experiment on some
slightly older nodes (2.6 GHz Pentium4 with Intel Corp. 82546EB
Gigabit Ethernet Controllers) show the performance that we would
actually expect:
Node A -> Node B: 115 MByte/s
Node A -> Node B -> Node C: 108 MByte/s

Are there any known performance problems with Intel 82541GI/PI or
82547GI Gigabit Ethernet cards?

Cheers,
Felix